講演情報
[PPS02-P14]Feature-Based Image Registration of Lunar Surface Images Under Varying Illumination Conditions
*中尾 康弘1、本田 親寿1 (1.会津大学)
A large number of lunar surface images have been acquired by lunar orbiters. These images are widely used for studies of impact craters, including the detection and classification of craters according to their degrees of degradation, which are important for understanding the history and geology of the lunar surface and for applications such as landing site evaluation. In recent years, crater detection and classification using machine learning have also been studied in lunar image analysis, because manual analysis of a large number of images covering wide areas requires a great amount of time and effort, and because machine learning based methods allow analyses to be performed in a more objective manner without relying on subjective human judgment. However, many of these studies are based on images taken under a single illumination condition. As illumination conditions affect the length of crater shadows and the visibility of ejecta, using multiple images of the same area taken under different illumination conditions is expected to enable analyses based on a wider variety of lunar surface features. In this study, an automatic image registration method for lunar surface image pairs taken under different illumination conditions was investigated. Lunar images acquired by the Narrow Angle Camera (NAC) onboard the Lunar Reconnaissance Orbiter (LRO) were used. The target area was a part of Mare Tranquillitatis, which contains craters with various degrees of degradation. Because a large number of lunar images exist and manual image registration requires significant time and effort, an automatic approach was used. Image pairs with different incidence angles and image pairs with opposite solar illumination directions were selected to examine the effects of illumination differences on image registration performance. A feature-based image registration method using the Scale-Invariant Feature Transform (SIFT) was applied, and a homography transformation was estimated to register the images. The results show that image pairs with moderate differences in incidence angle can be successfully registered with enough accuracy. In contrast, direct image registration becomes difficult when the difference in incidence angle is large, because shadow length and crater appearance change significantly. In such cases, a stepwise image registration approach using an intermediate image with a moderate incidence angle was effective. In addition, image pairs with opposite solar illumination directions showed lower registration performance even when incidence angles were similar. In this study, this issue was improved by applying Sobel filtering as a preprocessing step, which reduced the effect of reversed shadow directions and enabled successful image registration. These results indicate that image registration performance depends on illumination conditions. Furthermore, image registration can be achieved in some cases by adjusting the method according to the illumination conditions. The proposed approach is expected to be useful for future analyses of lunar surface images.